MinkowskiPE: Minkowski Positional Encoding for Spatiotemporal Perception
The paper adds a spacetime-aware positional encoding that keeps attention translation-invariant without changing the dot-product interface.
MinkowskiPE parameterizes Lorentz transformations with joint temporal and spatial coordinates, applying them to query and key features. The attention score then depends on relative spacetime displacement rather than absolute position. The authors report best results across nine molecular dynamics evaluations, plus a 9.9% KTH video-prediction MSE reduction versus the best baseline with about one-tenth the parameters. HF Daily Papers' note
MinkowskiPE parameterizes Lorentz transformations with joint temporal and spatial coordinates, applying them to query and key features. The attention score then depends on relative spacetime displacement rather than absolute position. The authors report best results across nine molecular dynamics evaluations, plus a 9.9% KTH video-prediction MSE reduction versus the best baseline with about one-tenth the parameters. HF Daily Papers' note
score 4